FusionCell: Cross-Attentive Fusion of Layout Geometry and Netlist Topology for Standard-Cell Performance Prediction

2026-05-22 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

FusionCell is a novel dual-modality predictor designed to accelerate standard-cell performance characterization by jointly analyzing routed layout geometry and netlist topology. This model employs a DeiT encoder to process three-layer routed layouts and a graph transformer to model heterogeneous device/net graphs. The key innovation lies in its topology-guided fusion mechanism, where the netlist acts as a structural "map" to query relevant physical regions in the layout, ensuring fine-grained spatial details are interpreted within their correct electrical context. Evaluated on a 7nm dataset built from the ASAP7 PDK, comprising over 19.5k cells across 149 types, FusionCell achieved an average Mean Absolute Percentage Error (MAPE) of 0.92% and superior Spearman/Kendall ranking correlation for six critical metrics: signal rise/fall delay, transition, and power. This approach delivers a 10^4x speedup over traditional circuit simulation.

Key takeaway

For Machine Learning Engineers tasked with accelerating standard-cell characterization, FusionCell provides a robust solution. You should consider integrating its topology-guided multimodal fusion to achieve a 10^4x speedup over traditional simulation while maintaining high accuracy (0.92% MAPE). This enables rapid design space exploration and efficient identification of Pareto-optimal cells, significantly reducing computational costs in your chip design process.

Key insights

Topology-guided fusion of layout geometry and netlist topology accurately predicts standard-cell performance.

Principles

• Layout R/C effects must be interpreted under correct electrical connectivity.
• Netlist topology defines the performance envelope.
• Layout geometry determines precise performance variation.

Method

FusionCell uses a DeiT encoder for multi-layer layouts and a graph transformer for heterogeneous netlists. These are fused via topology-guided graph-query/image-key cross-attention, then aggregated for MLP regression.

In practice

• Use DeiT for multi-layer routed layout encoding.
• Model netlists with heterogeneous device-net graph transformers.
• Implement topology-guided cross-attention for fusion.

Topics

• Standard Cell Characterization
• Multimodal Fusion
• Layout Geometry
• Netlist Topology
• Graph Transformers
• DeiT Encoder

Code references

• zhywhite/PreCell

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• SQuadGen: Generating Simple Quad Layouts via Chart Distance Fields — SQuadGen uses diffusion models and Chart Distance Fields to generate simple, artist-friendly quad layouts for 3D shapes.
• Topology-Aware Skeleton Detection via Lighthouse-Guided Structured Inference — Lighthouse-Skel proposed a dual-branch, topology-aware skeleton detection method using structured inference and a lighthouse-guided completion strategy.
• How Does AI Learn to See in 3D and Understand Space? — Bridging 2D AI predictions with 3D geometry via geometric fusion enables scalable, semantic 3D scene understanding.
• TOPCELL: Topology Optimization of Standard Cell via LLMs — TOPCELL uses LLMs and GRPO to optimize transistor topology, achieving significant speedup and quality in standard cell design.
• Rethinking the Pointer Loss in Table Structure Recognition: Geometry-Aware Pointer Loss for Spatial Locality — Geometry-Aware Pointer (GAP) Loss improves Table Structure Recognition by reweighting cross-entropy based on spatial proximity, focusing on adjacent-cell errors.
• Geometric Monomial (GEM): a family of rational 2N-differentiable activation functions — GEM activation functions offer $C^{2N}$-smoothness and rational arithmetic, outperforming GELU in various deep learning benchmarks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.